New procedure for estimating prediction certainty in decision trees

Decision trees estimate prediction certainty using the class distribution in the leaf responsible for the prediction. We introduce an alternative method that yields better estimates. For each instance to be predicted, our method inserts the instance to be classified in the training set with one of the possible labels for the target attribute; this procedure is repeated for each one of the labels. Then, by comparing the outcome of the different trees, the method can identify instances that might present some difficulties to be correctly classified, and attribute some uncertainty to their prediction. We perform an extensive evaluation of the proposed method, and show that it is particularly suitable for ranking and reliability estimations. The ideas investigated in this paper may also be applied to other machine learning techniques, as well as combined with other methods for prediction certainty estimation.